Jerry Liu

Jerry Liu is the co-founder and chief executive of LlamaIndex. He created GPT Index, the project that became LlamaIndex, and helped develop it from an open-source retrieval framework into infrastructure for document parsing, extraction, and automation. His work addresses a persistent problem for AI applications: a model can read a document’s words while missing the relationships that make them meaningful.

From machine learning to LlamaIndex

Jerry graduated summa cum laude from Princeton in 2017 with a computer science degree and a certificate in finance. He worked on Quora’s feed-ranking system, then joined Uber as an AI resident in 2018 before becoming a research scientist in 2019. At Uber ATG, his research covered learning-based compression and joint prediction and planning for autonomous driving. His coauthored publications include Deep Structured Reactive Planning and work on stereo-image, video, and LiDAR compression. He subsequently became a machine-learning engineering manager at Robust Intelligence, leading work on model monitoring and AI Firewall.

In November 2022, Jerry created GPT Index to help GPT-3 navigate information outside its training data. The project expanded into a framework with connectors, indexing, retrieval, and answer-synthesis components that developers could combine around their own data. In 2023, he and Simon Suo announced a company around LlamaIndex and an $8.5 million seed round led by Greylock.

The document became the problem

LlamaIndex made it easy to assemble a retrieval-augmented generation application: load information, retrieve relevant passages, and ask a model to compose an answer. But users encountered inaccurate answers, overwhelming tuning choices, changing source data, and PDFs whose formatting resisted ordinary text extraction.

Jerry’s 2024 introduction of managed parsing, ingestion, and retrieval explained a specific bottleneck. The team had developed methods for hierarchically indexing tables and text, but those methods depended on correctly recovering the tables first. LlamaParse produced structured Markdown that could feed the retrieval pipeline. Evaluations using Apple and Uber financial filings tested whether the resulting systems could answer questions about table cells.

His Caltrain timetable example makes the failure easy to understand. A parser can preserve the words and numbers while losing which departure time belongs to which column. The model then receives readable text with damaged relationships and can return the wrong train time. Jerry’s approach combines file-aware parsing, visual interpretation, and checking to reconstruct those relationships. As he explains in his 2026 document-context talk, a PDF describes how to draw a page; its tables and reading order must often be inferred from positions rather than recovered as explicit structures.

The company’s development focus has shifted toward LlamaParse, the local open-source parser LiteParse, and open parsing and extraction benchmarks. The original orchestration framework remains available. Jerry’s case for developer-led enterprise AI complements that shift: managed infrastructure should reduce data preparation work while leaving developers able to adapt applications to changing models, information, and business logic.

Retrieval decisions that depend on the task

Jerry’s technical guidance distinguishes problems that can look identical in a chatbot’s final answer:

• Measure retrieval and generation separately. His approach to production-ready RAG begins with a task-specific benchmark. Failing to retrieve the relevant passage requires a different fix from generating an incorrect answer from the right passage. He recommends testing parsing, chunk sizes, hybrid search, and metadata filters before adding more costly agent architectures. For a question about 2021 risk factors, a year filter can exclude semantically similar material from another reporting period.

• Search narrowly, then supply broader context. With small-to-big retrieval, the system searches precise passages and expands their surrounding context for answer synthesis. A large chunk’s embedding may obscure a specific fact among unrelated material, while a small passage alone may omit information needed to interpret it. Jerry’s long-context work extends this separation to passages or summaries linked to entire documents.

Jerry welcomed larger context windows as a way to simplify chunking, summarization, and comparisons within documents. He also retained a role for selective retrieval across collections too large to load on every request. A narrow lookup and an analysis spanning several reports need different amounts of context, with different costs and latency. By 2026, his account of agent-based retrieval placed more search decisions inside the agent: it can reformulate queries and inspect additional material as the task develops, reducing the need to encode every retrieval decision in a fixed pipeline.

From answering questions to doing document work

Jerry’s document-agent approach extends retrieval into tools for structured queries, document manipulation, enterprise connectors, and synchronized permissions. He distinguishes assistants that people guide interactively from background automations with more constrained steps and review before results move downstream. The distinction affects both the interface and how much control flow developers specify in advance.

The spreadsheet-agent introduction, which Jerry coauthored with Zhaoqi Li and Daniel Bustamante Ospina, illustrates the team’s approach. Merged cells, whitespace, and formatting can encode relationships that disappear in a CSV conversion. Their system first builds a semantic map of the sheet, then equips agents with specialized calculation and transformation tools. It supports normalization into two-dimensional tables and answers traceable to source cells.

His earlier Llama Agents architecture addressed deployment: specialists became reusable services communicating through a message queue, with either explicitly defined orchestration or LLM-directed delegation. Jerry presented it as an alpha in 2024, and described its internal use as constrained, simple RAG pipelines rather than general autonomous task solving.

Spend deeper reasoning selectively

Jerry advocates two-pass document processing. An inexpensive first pass makes a collection searchable; deeper visual processing then examines relevant pages, including difficult tables and charts. This avoids applying slow, costly visual interpretation to every page while preserving a way to inspect the details that matter. His LiteParse samples demonstrate local, model-free parsing through side-by-side comparisons and keyword matches highlighted on original document pages.

He also argues for model routing within the agent’s orchestration. Routing there can use task context and the logic of the workflow to choose models before execution and during the agent loop. Optimizing isolated model completions misses how the model mixture and orchestration work together.

On September 9, 2026, Jerry announced LlamaParse’s native ChatGPT connector, following the company’s Claude connector. It exposes parsing and extraction endpoints so the assistant can handle higher-level reasoning while specialized tools prepare document context and extract fields with grounding and citations.